pith. sign in

arxiv: 2606.20765 · v1 · pith:UXZ46XXHnew · submitted 2026-06-18 · 📡 eess.IV · cs.LG· q-bio.QM

Dataset-Aware Cold-Start Active Learning for Annotation-Efficient 3D Medical Image Segmentation

R\'emi Hattat , Marine Beaumont , Charline Bertholdt , Gabriela Hossu , Olivier Morel , Bailiang Chen This is my paper

Pith reviewed 2026-06-26 15:30 UTC · model grok-4.3

classification 📡 eess.IV cs.LGq-bio.QM
keywords cold-start active learning3D medical image segmentationself-supervised signalsannotation efficiencydataset-aware selectionDifficulty-Coverage Ratio
0
0 comments X

The pith

Dataset-aware cold-start active learning selects first 3D medical volumes by balancing unlabeled representativeness and difficulty signals.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper targets the cold-start problem in active learning for 3D medical image segmentation, where no labeled examples exist at the outset. It introduces a framework that draws two label-free signals from the unlabeled pool alone: local typicality in embedding space to capture representativeness, and reconstruction uncertainty to proxy difficulty. These signals are fused via a weighted geometric score whose balance is set by a closed-form pacing rule driven by the pool-wide Difficulty-Coverage Ratio. Experiments across four volumetric benchmarks show competitive downstream segmentation performance, with the largest improvements appearing when annotation budgets remain low to moderate.

Core claim

CSCS adapts initial sample selection to the geometry of an unlabeled dataset by combining local typicality and reconstruction-based uncertainty through a Difficulty-Coverage Ratio and closed-form pacing rule, producing competitive segmentation results with nnU-Net on BraTS, FeTA, Spleen, and an in-house fetal MRI dataset, especially under limited annotation budgets.

What carries the argument

The Difficulty-Coverage Ratio, a pool-level statistic that quantifies alignment between difficulty and representativeness, which sets the weighting in a geometric score for selecting the first annotation volumes.

If this is right

  • CSCS delivers competitive performance across four 3D medical segmentation datasets under varying annotation budgets.
  • Gains are largest in low-to-mid annotation regimes.
  • Sample selection is automatically adjusted to the unlabeled pool's internal structure.
  • The method eliminates the need for an initial labeled seed set before active learning begins.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same unlabeled signals could support cold-start selection in other volumetric imaging domains where annotation cost is high.
  • The closed-form pacing rule may reduce the need for dataset-specific tuning when moving between clinical sites.
  • If the two signals stay complementary across modalities, they could anchor broader self-supervised active learning pipelines.

Load-bearing premise

The two self-supervised signals remain informative and complementary proxies for representativeness and difficulty even before any task-specific labels exist.

What would settle it

If CSCS yields lower segmentation accuracy than random selection on any of the four benchmarks at low annotation budgets, the central claim would be contradicted.

Figures

Figures reproduced from arXiv: 2606.20765 by Bailiang Chen, Charline Bertholdt, Gabriela Hossu, Marine Beaumont, Olivier Morel, R\'emi Hattat.

Figure 1
Figure 1. Figure 1: Overview of the proposed CSCS pipeline. Self-supervised learning (SSL) pre￾training is performed on the unlabeled pool to extract, for each volume Xi, an embed￾ding zi, a typicality score T(Xi), and a reconstruction-based uncertainty score U(Xi). After diversity-preserving clustering into B groups, one sample per cluster is selected by maximizing the composite score S(Xi) = T(Xi) 1−γ ·U(Xi) γ . The pacing … view at source ↗
Figure 2
Figure 2. Figure 2: Segmentation performance as a function of annotation budget across four benchmarks. Top rows: mean Dice (%) and HD95 (mm) for BraTS, Spleen, and FeTA. Bottom row: DIANE dataset. Error bars: ±1 standard deviation over 3 independent runs. near-zero DCR (BraTS) keeps selection balanced, whereas higher positive DCR (DIANE) tilts it toward difficulty-aware selection. Cross-dataset consistency and statistical an… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative segmentation examples on the BraTS, FeTA and DIANE datasets at 20% annotation budget (B=77, B=8, and B=5, respectively). the strongest fixed baselines while being significantly more reliable than the weaker ones, using a single rule rather than one that happens to suit a given dataset. The boundary metric is consistent with this picture: HD95 ( [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
read the original abstract

Deep learning for 3D medical image segmentation requires extensive manual annotations, a major bottleneck in volumetric medical imaging. Active learning aims to reduce this burden by selecting informative samples for annotation, but most methods assume that an initial labeled set is already available. This leaves the cold-start problem largely unresolved: how to select the first volumes from a fully unlabeled pool before any task-specific model is trained. We propose CSCS, a Curriculum-Stratified Cold-Start framework that adapts initial sample selection to the structure of the unlabeled dataset. CSCS combines two self-supervised, label-free signals: local typicality, measuring representativeness in the embedding space, and reconstruction-based uncertainty, used as a proxy for sample difficulty. These signals are combined through a weighted geometric score, where the weighting is determined by a closed-form pacing rule based on the effective annotation budget and the Difficulty-Coverage Ratio, a pool-level statistic measuring the alignment between difficulty and representativeness. We evaluate CSCS on four 3D medical image segmentation benchmarks: BraTS, FeTA, Spleen, and an in-house fetal MRI dataset. Using nnU-Net as downstream segmentation model, CSCS shows consistently competitive performance across datasets and annotation budgets, with the strongest gains in low-to-mid annotation regimes. These results suggest that dataset-aware cold-start initialization can improve the robustness of active learning for 3D medical image segmentation by adapting sample selection to the geometry of the unlabeled pool.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes CSCS, a Curriculum-Stratified Cold-Start active learning framework for 3D medical image segmentation. It combines two self-supervised, label-free signals—local typicality in embedding space (representativeness) and reconstruction-based uncertainty (difficulty proxy)—via a weighted geometric score whose weighting is set by a closed-form pacing rule driven by the Difficulty-Coverage Ratio, a pool-level statistic. The method is evaluated on four benchmarks (BraTS, FeTA, Spleen, in-house fetal MRI) with nnU-Net as the downstream model and claims consistently competitive performance, strongest in low-to-mid annotation regimes.

Significance. If the performance claims hold, the work addresses a practically important gap in annotation-efficient 3D medical segmentation by providing a dataset-aware cold-start solution that avoids reliance on an initial labeled set. The label-free signals and closed-form pacing rule are potential strengths if they prove robust and non-circular; reproducible code or machine-checked derivations would further strengthen the contribution.

major comments (2)
  1. [Abstract] Abstract: the claim of competitive results on four benchmarks supplies no implementation details, baseline comparisons, statistical tests, or ablation of the pacing rule, so the evaluation claims cannot be verified from the given text.
  2. [Abstract] Abstract: the pacing rule is described as closed-form and derived from the Difficulty-Coverage Ratio, but without the full equations it is impossible to confirm whether the ratio or weighting reduces to a fitted quantity or introduces circular dependence on the selection itself.
minor comments (1)
  1. [Abstract] Abstract: the term 'Curriculum-Stratified' is introduced without a brief definition or link to how stratification interacts with the two signals.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the abstract should better support its claims with pointers to evaluation details and the pacing rule derivation. We address the two major comments below and will revise the abstract in the next version.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of competitive results on four benchmarks supplies no implementation details, baseline comparisons, statistical tests, or ablation of the pacing rule, so the evaluation claims cannot be verified from the given text.

    Authors: The abstract is necessarily concise. The full manuscript provides implementation details and nnU-Net usage in Section 3, baseline comparisons and quantitative results across the four datasets in Section 4 (Tables 1-4, Figures 2-5), statistical significance tests in Section 4.3, and an ablation of the pacing rule in Section 4.4. We will revise the abstract to add a brief clause referencing the evaluation protocol, annotation budgets, and ablation studies. revision: yes

  2. Referee: [Abstract] Abstract: the pacing rule is described as closed-form and derived from the Difficulty-Coverage Ratio, but without the full equations it is impossible to confirm whether the ratio or weighting reduces to a fitted quantity or introduces circular dependence on the selection itself.

    Authors: The Difficulty-Coverage Ratio is defined as a pool-level statistic (Equation 3, Section 3.2) computed exclusively from the unlabeled data prior to any selection. The pacing rule (Equation 4) is a closed-form expression depending only on this ratio and the target annotation budget; it contains no learned parameters and introduces no dependence on the chosen subset. We will revise the abstract to include the key equations or an explicit statement that the ratio is pre-computed and dataset-level. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The described framework uses two self-supervised label-free signals combined via a closed-form pacing rule driven by a pool-level Difficulty-Coverage Ratio statistic. No equations, self-citations, or fitted parameters are shown to reduce the central claims to tautological inputs by construction. The derivation remains independent of the target selection outcomes and relies on external empirical benchmarks for validation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method rests on two domain assumptions about the meaning of self-supervised signals and on the effectiveness of a closed-form pacing rule derived from a single pool statistic; no free parameters or new entities are explicitly introduced in the abstract.

axioms (2)
  • domain assumption Local typicality in embedding space measures representativeness of unlabeled volumes
    Used as one of the two label-free signals for initial sample selection.
  • domain assumption Reconstruction error serves as a valid proxy for sample difficulty
    Used as the second label-free signal combined with typicality.

pith-pipeline@v0.9.1-grok · 5815 in / 1217 out tokens · 30185 ms · 2026-06-26T15:30:33.228270+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Nature Methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods18(2), 203–211 (2021)

    2021

  2. [2]

    In: BrainLes@MICCAI

    Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images. In: BrainLes@MICCAI. pp. 272–284 (2022) CSCS for 3D medical image segmentation 19

    2022

  3. [3]

    Medical Image Analysis63, 101693 (2020)

    Tajbakhsh, N., Jeyaseelan, L., Li, Q., Chiang, J.N., Wu, Z., Ding, X.: Embracing imperfect datasets: A review of deep learning solutions for medical image segmen- tation. Medical Image Analysis63, 101693 (2020)

    2020

  4. [4]

    Nature Communications12(1), 5915 (2021)

    Wang, S., Li, C., Wang, R., Liu, Z., et al.: Annotation-efficient deep learning for au- tomatic medical image segmentation. Nature Communications12(1), 5915 (2021)

    2021

  5. [5]

    Medical Image Analysis95, 103201 (2024)

    Wang, H., Jin, Q., Li, S., Liu, S., Wang, M., Song, Z.: A comprehensive survey on deep active learning in medical image analysis. Medical Image Analysis95, 103201 (2024)

    2024

  6. [6]

    Computer Sciences Technical Report 1648, University of Wisconsin-Madison (2009)

    Settles, B.: Active learning literature survey. Computer Sciences Technical Report 1648, University of Wisconsin-Madison (2009)

    2009

  7. [7]

    In: ICLR (2018)

    Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core- set approach. In: ICLR (2018)

    2018

  8. [8]

    In: ICML

    Gal, Y., Islam, R., Ghahramani, Z.: Deep Bayesian active learning with image data. In: ICML. pp. 1183–1192 (2017)

    2017

  9. [9]

    In: ICLR (2020)

    Ash, J.T., Zhang, C., Krishnamurthy, A., Langford, J., Agarwal, A.: BADGE: Batch active learning by diverse gradient embeddings. In: ICLR (2020)

    2020

  10. [10]

    In: NeurIPS (2022)

    Yehuda, O., Dekel, A., Hacohen, G., Weinshall, D.: Active learning through a covering lens. In: NeurIPS (2022)

    2022

  11. [11]

    In: NeurIPS

    Huang, S.J., Jin, R., Zhou, Z.H.: Active learning by querying informative and representative examples. In: NeurIPS. pp. 892–900 (2010)

    2010

  12. [12]

    In: ICCV

    Sinha, S., Ebrahimi, S., Darrell, T.: Variational adversarial active learning. In: ICCV. pp. 5972–5981 (2019)

    2019

  13. [13]

    In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T

    Wang, X., Lian, L., Yu, S.X.: Unsupervised Selective Labeling for More Effective Semi-Supervised Learning. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision – ECCV 2022. LNCS, vol. 13690, pp. 427–445. Springer, Cham (2022)

    2022

  14. [14]

    In: MICCAI

    Liu, H., Li, H., Yao, X., Fan, Y., et al.: COLosSAL: A benchmark for cold-start active learning for 3D medical image segmentation. In: MICCAI. pp. 25–34 (2023)

    2023

  15. [15]

    In: Medical Imaging with Deep Learning

    Chen, L., Bai, Y., Huang, S., Lu, Y., Wen, B., Yuille, A.L., Zhou, Z.: Making Your First Choice: To Address Cold Start Problem in Medical Active Learning. In: Medical Imaging with Deep Learning. Proceedings of Machine Learning Research, vol. 227, pp. 496–525 (2024)

    2024

  16. [16]

    In: MICCAI (2025)

    Zhu, N., Ye, P., Zhong, L., Yue, Q., Zhang, S., Wang, G.: CSAL-3D: Cold-start active learning for 3D medical image segmentation via SSL-driven uncertainty- reinforced diversity sampling. In: MICCAI (2025)

    2025

  17. [17]

    In: ICML

    Hacohen, G., Dekel, A., Weinshall, D.: Active learning on a budget: Opposite strategies suit high and low budgets. In: ICML. pp. 8175–8195 (2022)

    2022

  18. [18]

    In: NeurIPS

    Hacohen, G., Weinshall, D.: How to select which active learning strategy is best suited for your specific problem and budget. In: NeurIPS. pp. 13395–13407 (2023)

    2023

  19. [19]

    In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024

    Ma, S., Du, H., Curran, K.M., Lawlor, A., Dong, R.: Adaptive Curriculum Query Strategy for Active Learning in Medical Image Classification. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024. pp. 48–57 (2024)

    2024

  20. [20]

    In: NeurIPS 2020 Workshop on Pre-registration in Machine Learning

    Chandra, A.L., Desai, S.V., Devaguptapu, C., Balasubramanian, V.N.: On Initial Pools for Deep Active Learning. In: NeurIPS 2020 Workshop on Pre-registration in Machine Learning. Proceedings of Machine Learning Research, vol. 148, pp. 14–32 (2021)

    2020

  21. [21]

    In: Medical Image Com- puting and Computer Assisted Intervention (MICCAI)

    Nath, V., Yang, D., Roth, H.R., Xu, D.: Warm-Start Active Learning with Proxy Labels and Selection via Semi-Supervised Fine-Tuning. In: Medical Image Com- puting and Computer Assisted Intervention (MICCAI). pp. 297–308 (2022)

    2022

  22. [22]

    arXiv preprint arXiv:2402.02561 (2024) 20 Hattat et al

    Yuan, J., et al.: Foundation Model Makes Clustering a Better Initialization for Cold-Start Active Learning. arXiv preprint arXiv:2402.02561 (2024) 20 Hattat et al

    work page arXiv 2024

  23. [23]

    Transactions on Machine Learning Research (2025)

    Lüth, C.T., Traub, J., Kahl, K.-C., Bungert, T.J., Klein, L., Krämer, L., Jaeger, P.F., Isensee, F., Maier-Hein, K.: nnActive: A Framework for Evaluation of Ac- tive Learning in 3D Biomedical Segmentation. Transactions on Machine Learning Research (2025)

    2025

  24. [24]

    Trans- actions on Machine Learning Research (2026)

    Lüth, C.T., Traub, J., Kahl, K.-C., Bungert, T.J., Klein, L., Krämer, L., Jaeger, P.F., Maier-Hein, K., Isensee, F.: Finally Outshining the Random Baseline: A Sim- ple and Effective Solution for Active Learning in 3D Biomedical Imaging. Trans- actions on Machine Learning Research (2026)

    2026

  25. [25]

    Nature Communications13(1), 4128 (2022)

    Antonelli, M., Reinke, A., Bakas, S., et al.: The Medical Segmentation Decathlon. Nature Communications13(1), 4128 (2022)

    2022

  26. [26]

    Scientific Data8(1), 167 (2021)

    Payette, K., de Dumast, P., Kebiri, H., Ezhov, I., Paetzold, J.C., Shit, S., et al.: An automatic multi-tissue human fetal brain segmentation benchmark using the fetal tissue annotation dataset. Scientific Data8(1), 167 (2021)

    2021

  27. [27]

    In: CVPR

    Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of Swin transformers for 3D medical image analysis. In: CVPR. pp. 20730–20740 (2022)

    2022

  28. [28]

    In: CVPR Workshops

    Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.A.: Performance prediction for semantic segmentation by a self-supervised image reconstruction decoder. In: CVPR Workshops. pp. 4394–4403 (2022)

    2022

  29. [29]

    In: Proc

    Arthur, D., Vassilvitskii, S.: k-means++: The Advantages of Careful Seeding. In: Proc. ACM-SIAM Symposium on Discrete Algorithms (SODA) (2007)

    2007

  30. [30]

    Conover, W.J.: Practical Nonparametric Statistics. 3rd edn. John Wiley & Sons (1999)

    1999

  31. [31]

    Kendall, A., Gal, Y.: What uncertainties do we need in Bayesian deep learn- ing for computer vision? In: Advances in Neural Information Processing Systems (NeurIPS). pp. 5574–5584 (2017)

    2017

  32. [32]

    arXiv preprint arXiv:2106.13731 (2021)

    Wright, L., Demeure, N.: Ranger21: A synergistic deep learning optimizer. arXiv preprint arXiv:2106.13731 (2021)

    work page arXiv 2021

  33. [33]

    In: Advances in Neural Information Process- ing Systems (NeurIPS) (2019)

    Paszke, A., Gross, S., Massa, F., et al.: PyTorch: An Imperative Style, High- Performance Deep Learning Library. In: Advances in Neural Information Process- ing Systems (NeurIPS) (2019)

    2019

  34. [34]

    MONAI: An open-source framework for deep learning in healthcare

    Cardoso, M.J., Li, W., Brown, R., et al.: MONAI: An Open-Source Framework for Deep Learning in Healthcare. arXiv preprint arXiv:2211.02701 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  35. [35]

    Nature Methods21, 195–212 (2024)

    Maier-Hein, L., Reinke, A., Godau, P., et al.: Metrics reloaded: Recommendations for image analysis validation. Nature Methods21, 195–212 (2024)

    2024

  36. [36]

    Knowledge-Based Systems241, 108278 (2022)

    Jin, Q., Yuan, M., Qiao, Q., Song, Z.: One-shot active learning for image seg- mentation via contrastive learning and diversity-based sampling. Knowledge-Based Systems241, 108278 (2022)

    2022

  37. [37]

    arXiv preprint arXiv:2601.18532 (2026)

    Levy, D., Assayag, B., Gaspar, L., Shimshoni, I., Specktor-Fadida, B.: From Cold Start to Active Learning: Embedding-Based Scan Selection for Medical Image Seg- mentation. arXiv preprint arXiv:2601.18532 (2026)

    work page arXiv 2026

  38. [38]

    arXiv preprint arXiv:2508.03441 (2025)

    Zhu, N., Ma, X., Zhang, S., Wang, G.: MedCAL-Bench: A Comprehensive Bench- mark on Cold-Start Active Learning with Foundation Models for Medical Image Analysis. arXiv preprint arXiv:2508.03441 (2025)

    work page arXiv 2025